LlaMario：基于大语言模型的可控马里奥关卡生成

doi:10.11996/JG.j.2095-302X.2026030511

摘要/Abstract

摘要：

在游戏开发过程中，程序内容生成(PCG)能够有效降低游戏开发成本，提高关卡设计的多样性。近年来，程序内容生成与机器学习的结合在游戏关卡生成方面取得了显著进展。针对现有方法在生成控制性和适应复杂设计意图方面仍存在局限性的问题，提出了LlaMario模型。通过将马里奥游戏关卡转换成字符矩阵并交由大语言模型(LLMs)训练，最终将生成的关卡字符再转换成可玩的马里奥游戏关卡。具体而言，采用LLMs Llama 3.1-8B-instruct，并结合高效语言模型微调框架Unsloth，使模型实现了更强的关卡理解与生成能力。通过构建4万条基于LLMs Gemini理解关卡数据的指令集，并采用数据高效的指令微调策略Alpaca进行关卡数据集的构建，使得经过训练后的LlaMario能够基于自然语言描述生成高可玩性的马里奥关卡。实验结果表明，在关卡逻辑合理性、可玩性、复杂性和生成质量方面表现优异，能够生成用户期望的马里奥游戏关卡。该方法框架具有通用性，可扩展至其他基于瓷砖地图的关卡生成任务。

关键词: 程序内容生成, 大语言模型, 游戏关卡生成, 指令微调, 可控关卡生成

Abstract:

In game development, Procedural Content Generation (PCG) effectively reduces development costs and enhances the diversity of level design. Recent advances in integrating PCG with machine learning have demonstrated significant progress in game level generation. However, existing methods still exhibit limitations in generation controllability and adaptability to complex design intentions. To address these limitations, the LlaMario model was proposed to convert Mario game levels into character matrices for training with Large Language Models (LLMs), and then transformed the generated symbolic sequences into playable Mario levels. Specifically, we adopted the Llama 3.1-8B-instruct model combined with the efficient language model to fine-tune the Unsloth framework, empowering the model with enhanced level comprehension and generation capabilities. An instruction set of 40 000 entries derived from the Gemini model’s comprehension of level data was constructed and employed together with the data-efficient Alpaca instruction-tuning strategy for dataset construction; the resulting LlaMario trained on this dataset generated highly playable Mario levels from natural language descriptions. Experimental results validated that the proposed model achieved exceptional performance in level logical coherence, playability, complexity, and generation quality, successfully producing user-intended Mario game levels. Furthermore, the proposed framework demonstrated generality and was extended to other tile-based level generation tasks.

Key words: procedural content generation, large language models, game level generation, instruction fine-tuning, controllable level generation

中图分类号:

TP317
TP18

耿钰轩, 卢奕南, 伍铁如, 李文辉, 马锐. LlaMario：基于大语言模型的可控马里奥关卡生成[J]. 图学学报, 2026, 47(3): 511-523.

GENG Yuxuan, LU Yinan, WU Tieru, LI Wenhui, MA Rui. LlaMario: controllable Mario level generation based on large language models[J]. Journal of Graphics, 2026, 47(3): 511-523.

图/表 14

图1 LlaMario模型架构图，使用关卡指令可控的生成马里奥关卡

Fig. 1 Architecture of the LlaMario model for instruction-controllable generation of Mario levels

表1 马里奥关卡字符-瓷砖贴图对照表

Table 1 Character-tile texture correspondence table for Mario levels

瓷砖类型	字符表示	场景表示
空气	‘-’
不可破坏障碍物	‘X’
可破坏障碍物	‘S’
问号砖块	‘?’/‘Q’
硬币	‘o’
敌人	‘E’
左管道头	‘<’
右管道头	‘>’
左管道体	‘[’
右管道体	‘]’
炮台头	‘B’
炮台体	‘b’

图2 数据集构建流程图，生成包含两类数据的多个马里奥关卡数据集

Fig. 2 Dataset construction workflow diagram for generating multiple Mario level datasets containing two categories of data

图3 AIPrompt_40K数据样例

Fig. 3 AIPrompt_40K data samples

图4 Mario管道样式实例

Fig. 4 Examples of Mario pipe styles

表2 40k_AIPrompt数据集重复率表

Table 2 Duplication rate table of the 40k_AIPrompt dataset

N值	NRR↓	NRR-cs↓	ΔNRR↓	ΔNRR-cs↓
3	0.999 999	0.999 910	0.000 002	0.000 042
4	0.999 997	0.999 875	0.000 012	0.000 236
5	0.999 995	0.999 841	0.000 048	0.000 680

表3 模型生成关卡可玩性

Table 3 Percentage of playable levels generated

关卡生成模型	关卡可玩性/%↑	A*与对应模型人工筛查差值
Llama 3.1-8B-instruct	0	一
LSTM^[25](A*)	31.0	一
MarioGPT^[7](A*)	84.0	一
LlaMario_NoAIPrompt_40K_e3(A*)	72.5	一
LlaMario_NoAIPrompt_200K_e1(A*)	84.0	一
LlaMario_AIPrompt_40K_e3(A*)	78.5	一
LlaMario_AIPrompt_40K_e5(A*)	88.5	一
Llama2-7B_AIPrompt_40K_e5(A*)	80.5	一
MarioGPT^[7](人工筛查)	88.0	4.0
LlaMario_NoAIPrompt_40K_e3 (人工筛查)	84.0	11.5
LlaMario_NoAIPrompt_200K_e1 (人工筛查)	90.0	6.0
LlaMario_AIPrompt_40K_e3 (人工筛查)	87.0	8.5
LlaMario_AIPrompt_40K_e5 (人工筛查)	94.0	5.5

图5 部分LlaMario模型生成的马里奥可玩关卡切片

Fig. 5 Slices of playable Mario levels generated by the LlaMario model

表4 可玩性随重试预算变化的统计

Table 4 Statistics of playability as a function of the retry budget

重试次数/次	关卡可玩性/%↑	丢弃关卡数/个↓
K=0	89.4	53
K=1	99.2	4
K=2	100.0	0

表5 模型生成关卡贴图准确性/%

Table 5 Tile texture accuracy of model-generated levels/%

关卡生成模型	关卡贴图准确性↑
MarioGPT	64.0
LlaMario模型	72.0
LlaMario模型+后处理	98.0

图6 模型生成关卡中影响准确性的问题示例

Fig. 6 Examples of issues affecting accuracy in model-generated levels

表6 关卡准确性评测结果表

Table 6 Level accuracy evaluation results

关卡生成模型	关卡准确性总分	字符合法率/%	结构一致性/%	管道一致性/%
MarioGPT	81.3	100.0	72.5	82.8	69.8
LlaMario模型	81.7	100.0	93.8	46.1	86.7
LlaMario模型+后处理	96.6	100.0	97.5	89.6	99.4

表7 模型生成关卡可控性/%

Table 7 Controllability of model-generated levels/%

关卡生成模型	关卡可控性↑
LlaMario_NoAIPrompt_40k_e3	23.0
LlaMario_NoAIPrompt_200k_e1	21.0
LlaMario_AIPrompt_40k_e3	76.0
LlaMario_AIPrompt_40k_e5	62.5

图7 基于输入引导的部分生成的马里奥关卡切片((a) 输入用户需求为 “我希望生成一个地面不平整的关卡，有一些障碍阻挡前进”；(b) 输入用户需求为“我希望生成一个适合新手游玩的简单关卡”；(c) 输入用户需求为“我希望生成一个有比较多管道的关卡”；(d) 输入用户需求为 “我希望生成一个需要很多跳跃且充满挑战性的关卡”)

Fig. 7 Slices of partially generated Mario levels based on input guidance ((a) The input user requirement is “I want to generate a level with uneven ground and some obstacles blocking progress”; (b) The input user requirement is “I want to generate a simple level suitable for new players”; (c) The input user requirement is “I want to generate a level with relatively many pipes”; (d) The input user requirement is “I want to generate a level that requires many jumps and is highly challenging”)

参考文献 27

[1]	SHAKER N, TOGELIUS J, NELSON M J. Procedural content generation in games[M]. Cham: Springer, 2016: 1-2.
[2]	RISI S, TOGELIUS J. Increasing generality in machine learning through procedural content generation[J]. Nature Machine Intelligence, 2020, 2(8): 428-436. DOI
[3]	GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial networks[J]. Communications of the ACM, 2020, 63(11): 139-144. DOI URL
[4]	GIACOMELLO E, LANZI P L, LOIACONO D. DOOM level generation using generative adversarial networks[C]//2018 IEEE Games, Entertainment, Media Conference. New York: IEEE Press, 2018: 316-323.
[5]	FONTAINE M C, LIU R L, KHALIFA A, et al. Illuminating mario scenes in the latent space of a generative adversarial network[EB/OL]. [2025-06-15]. https://aaai.org/papers/05922-illuminating-mario-scenes-in-the-latent-space-of-a-generative-adversarial-network/.
[6]	SUMMERVILLE A, SNODGRASS S, GUZDIAL M, et al. Procedural content generation via machine learning (PCGML)[J]. IEEE Transactions on Games, 2018, 10(3): 257-270. DOI URL
[7]	SUDHAKARAN S, GONZÁLEZ-DUQUE M, FREIBERGER M, et al. MarioGPT: open-ended text2level generation through large language models[C]//The 37th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2023: 2361.
[8]	SUMMERVILLE A J, SNODGRASS S, MATEAS M, et al. The VGLC: the video game level corpus[EB/OL]. [2025-06-15]. https://arxiv.org/abs/1606.07487.
[9]	TOUVRON H, LAVRIL T, IZACARD G, et al. LLaMA: open and efficient foundation language models[EB/OL]. [2025-06-15]. https://arxiv.org/abs/2302.13971.
[10]	RADFORD A, WU J, CHILD R, et al. Language models are unsupervised multitask learners[R]. San Francisco: OpenAI, 2019.
[11]	HAN D, HAN M, Unsloth Team. Unsloth[EB/OL]. [2025-02-21]. http://github.com/unslothai/unsloth.
[12]	BROWNE C, MAIRE F. Evolutionary game design[J]. IEEE Transactions on Computational Intelligence and AI in Games, 2010, 2(1): 1-16.
[13]	SMITH A M, MATEAS M. Answer set programming for procedural content generation: a design space approach[J]. IEEE Transactions on Computational Intelligence and AI in Games, 2011, 3(3): 187-200. DOI URL
[14]	LIU J L, SNODGRASS S, KHALIFA A, et al. Deep learning for procedural content generation[J]. Neural Computing and Applications, 2021, 33(1): 19-37. DOI
[15]	SUMMERVILLE A, MATEAS M. Super mario as a string: platformer level generation via LSTMs[EB/OL]. [2025-06-15]. https://doi.org/10.26503/dl.v2016i1.752.
[16]	BEUKMAN M, CLEGHORN C W, JAMES S. Procedural content generation using neuroevolution and novelty search for diverse video game levels[C]//The Genetic and Evolutionary Computation Conference. New York: ACM, 2022: 1028-1037.
[17]	VOLZ V, SCHRUM J, LIU J L, et al. Evolving mario levels in the latent space of a deep convolutional generative adversarial network[C]//The Genetic and Evolutionary Computation Conference. New York: ACM, 2018: 221-228.
[18]	GONZÁLEZ-DUQUE M, PALM R B, HAUBERG S, et al. Mario plays on a manifold: generating functional content in latent space through differential geometry[C]//2022 IEEE Conference on Games. New York: IEEE Press, 2022: 385-392.
[19]	SARKAR A, COOPER S. Generating and blending game levels via quality-diversity in the latent space of a variational autoencoder[C]//The 16th International Conference on the Foundations of Digital Games. New York: ACM, 2021: 3.
[20]	MOURET J B, CLUNE J. Illuminating search spaces by mapping elites[EB/OL]. [2025-06-15]. https://arxiv.org/abs/1504.04909.
[21]	TODD G, EARLE S, NASIR M U, et al. Level generation through large language models[C]//The 18th International Conference on the Foundations of Digital Games. New York: ACM, 2023: 70.
[22]	Gemini Team Google. Gemini: a family of highly capable multimodal models[EB/OL]. [2025-06-15]. https://arxiv.org/abs/2312.11805.
[23]	TAORI R, GULRAJANI I, ZHANG T Y, et al. Alpaca: a strong, replicable instruction-following model[R]. Stanford: Stanford Center for Research on Foundation Models, 2023.
[24]	SHINN N, CASSANO F, GOPINATH A, et al. Reflexion: language agents with verbal reinforcement learning[C]//The 37th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2023: 377.
[25]	Amidos2006. Mario-AI-Framework[EB/OL]. [2025-02-21]. https://github.com/amidos2006/Mario-AI-Framework.
[26]	TOGELIUS J, KARAKOVSKIY S, BAUMGARTEN R. The 2009 Mario AI competition[C]//IEEE Congress on Evolutionary Computation. New York: IEEE Press, 2010: 1-8.
[27]	LI Y H, WANG H R, ZHANG C. Assessing logical puzzle solving in large language models: insights from a minesweeper case study[EB/OL]. [2025-06-15]. https://aclanthology.org/2024.naacl-long.4/.